Prediction by Posterior Estimation in Virtual Screening

Abstract

The ability to rank molecules according to their effectiveness in some domain, e.g. pesticide, drug, is important due to the cost of synthesizing and testing chemical compounds. Virtual screening seeks to do this computationally with potential savings of millions of pounds and large profits associated with reduced time to market. A current leading machine learning algorithm in this area – Binary kernel discrimination produces scores based on the estimated likelihood ratio of active to inactive compounds that are then ranked. As the prediction by posterior estimation is effective for noisy high dimensional data, this paper aims to estimate, directly the posterior probability of molecules being active and rank the molecules based on that probability instead. This can be done by using a non-parametric logistic or probit regression. The complexity is controlled by penalizing the likelihood function. The spare logistic regression is extended to accommodate more sparse solutions. The 11 activity classes from the MDL Drug Data Report (MDDR) database are used. The results are found to be less accurate to a currently leading approach but are still comparable in a number of cases.